AI tools are increasingly present in the exam room — but a handful of startups want AI to replace doctors.

getty

Most clinical AI companies today focus on deploying products for functions focused on making doctors faster such as reducing their burnout or helping them make more informed decisions. For example, AI scribes like Abridge and Cleo reduce documentation time and medical-focused AIs like OpenEvidence help doctors make more informed decisions.

But a smaller group of startups is pushing a different narrative: they’re not building copilots. They’re trying to replace physicians. Here are three of those companies.

Doctronic: the AI doctor with prescribing power

Doctronic, a New York-based startup, offers a HIPAA-compliant platform it describes as a free, 24/7 personal AI doctor that can make clinical decisions on its own. The company raised $40 million in a Series B round in March, co-led by Abstract and Lightspeed Venture Partners, bringing its total funding to $65 million.

Doctronic is working in the state of Utah which is allowing it to legally renew prescriptions for patients managing chronic conditions without a doctor approving the prescription. It's the first time a state has authorized an AI platform to handle routine prescribing on its own.

Six months in, the data give show that the AI approved renewals without any human involvement 72% of the time for 192 drugs treating conditions like hypertension, diabetes and depression. In the other 28% of cases, the system escalated to a human physician on its own, typically because it needed an additional step like needing labs or it flagged a complication. Independent physician reviewers agreed the escalation was the right call 69% of the time.

MORE FOR YOU

Where the AI did grant a renewal without escalating, physicians agreed with its decision 91% of the time. Yet in 3% of cases, reviewing physicians disagreed with the AI’s call entirely.

As of this writing, Utah’s Office of AI Policy says no serious safety incidents have been reported. Still, the Utah Medical Licensing Board and the American Medical Association have both raised accountability concerns, and researchers have flagged security vulnerabilities in the pilot that regulators are still working through.

Certuma: an AI replacement for urgent care

Certuma was founded by serial entrepreneur Martin Varsavsky and is targeting common ailments that typically fill up urgent care waiting rooms. The Austin-based company has raised$10 million in seed funding, led by Joe Lonsdale’s 8VC, at a $60 million valuation.

Certuma is trying to be the first AI doctor with formal FDA approval. Its initial target list is 25 low-risk, high-frequency conditions, including UTIs, strep throat and sinusitis. They were chosen specifically because the clinical decision trees are well established and low-variance.

Certuma is also running a two-track strategy: a slower, higher-bar path through the FDA in the U.S., and a faster international track in Varsavsky’s native Argentina. There, the company has already worked with Argentinian regulators, aided by a deregulation-friendly government, to offer diagnosis and prescribing advice through a consumer-facing chatbot.

It’s a longer-term vision is to pair the AI with autonomous robotic booths where patients could get imaging or blood draws done.

Ada Health: AI triage without a waiting room

Ada Health has taken a related approach, building an AI health companion that helps people understand their symptoms and decide what to do next — often without ever routing them to a doctor.

The Berlin-based company has raised $242 million. Their most significant financing was a $120 million Series B round backed by investors including Inteligo Bank, Farallon Capital, and Red River West.

Ada Health says says it has 13 million users worldwide who have completed more than 32 million symptom assessments.

Ada points to a peer-reviewed BMJ study it led comparing eight symptom-checker apps on coverage, accuracy and safety where its own tool came out on top. It’s less "AI doctor" than "AI front door," but the effect is the same: fewer visits that require an actual physician.

Bigger AI players are building copilots, not replacements

It’s worth contrasting all of this with what the major AI companies are doing. Anthropic launched Claude for Healthcare in January 2026, at the J.P. Morgan Healthcare Conference, framed as a tool to support physicians. That’s also been the position of most large health systems where the goal is for AI to handle documentation, triage support and administrative drudgery, freeing doctors to spend more time on the judgment calls only they can make.

But even that line is getting harder to hold at the edges. Revere Health has already made some of the first AI-linked workforce cuts in U.S. healthcare with cuts to medical coding, billing and documentation staff, not physicians. Revere eliminated roughly 200 jobs, about 7% of its workforce. But watching AI quietly hollow out the roles around physicians, even while doctors themselves stay untouched, may be a preview of where the pressure could head next.

And physicians themselves are uneasy: in a recent Sermo poll, 58% said they believe AI will either diminish their role or make doctors obsolete outright.

Here’s the regulatory patchwork that's making this possible

None of these companies can just decide to replace a doctor. Two separate regulatory systems have to change for that to happen: federal device approval and state medical licensing. Right now, both are being tested at once, in different ways, by different companies.

On the federal side, the FDA has cleared more than 1,000 AI/ML-enabled medical devices. But the overwhelming majority (about 95% to 97%) went through the 510(k) pathway. This only requires a device to be "substantially equivalent" to something already on the market. That pathway isn’t built for autonomous treatment decisions.

The De Novo pathway, used for novel low-to-moderate-risk devices, has produced exactly one landmark precedent for autonomy: LumineticsCore (formerly IDx-DR), authorized back in 2018 as the first fully autonomous AI diagnostic system, for detecting diabetic retinopathy from a retinal scan with no physician interpretation required.

That’s the model Certuma is explicitly trying to replicate for its 25 target conditions. As of now, though, no autonomous AI prescribing service has been cleared through any FDA pathway.

On the state side, the path being used is entirely different, and arguably more consequential in the near term. Utah didn’t get Doctronic through a licensing exception — the company isn’t licensed to "practice medicine" at all. Instead, Utah’s Office of AI Policy, operating under an AI regulatory sandbox law passed in 2024, signed a formal non-enforcement agreement. The state agreed not to prosecute Doctronic under its unprofessional-conduct and unlicensed-practice statutes in exchange for the company adhering to a contract with specific safety and privacy guardrails. It only covers renewals of existing prescriptions, not new diagnoses or first-time prescribing.

Yet that carve-out is running directly into a wave of opposite-direction legislation elsewhere. Oregon and Delaware now explicitly bar any "nonhuman entity" from using a licensed clinical title — physician, nurse, PA — or being credentialed to practice at all.

By mid-2026, roughly 37 states had passed or introduced AI-in-healthcare legislation, and a common thread running through laws in California, Texas, Nebraska and elsewhere is a requirement that a licensed human clinician must make the final call on anything resembling a medical necessity determination. Utah's sandbox is very much the exception, and it's already facing pushback from the state's own Medical Licensing Board and the AMA.

Companies like Doctronic and Certuma are making a clear argument: if an algorithm can safely handle the routine 80%, human physicians can focus on the hard 20%, and patients get faster access either way.

But here’s the counterargument: medicine isn’t just pattern-matching. A sore throat can be strep, or it can be the first sign of something a good clinician would catch and an algorithm might miss.

The Utah prescription deal, the FDA filings, the malpractice questions nobody’s fully answered yet — all of that is where this story is actually going to be decided.

Ultimately, the safest bet is still that AI reshapes medicine rather than replaces the people who practice it. But for the first time, there are venture-backed companies, state governments and regulators actively testing what it would look like if that safest bet turned out to be wrong.